Statistical Downscaling: Predictor Selection and Model Evaluation for Future Rainfall Projection in Hawai‛i

Abstract:

Statistical downscaling methods bridge the gap between global-scale climate changes represented in general circulation models (GCMs) and local-scale regional impacts that are most relevant for decision makers. While traditional statistical methods such as multiple linear regression (MLR) are still useful tools for statistical downscaling, finding the best predictive large-scale climate information for the targeted local climate variables remains a challenge. A new method for large-scale atmospheric predictor evaluation is presented for use in statistical downscaling to project rainfall for the Hawaiian Islands. A pool of sixteen commonly used predictor variables (e.g. geopotential height, moisture transport, vertical temperature gradient) were assessed for correlation with rainfall and multicollinearity among other predictors. Linear regression was used to derive relationships between all possible predictor variable combinations and seasonal rainfall, and models having a delta Akaike Information Criterion (AIC) value of 2 or lower were retained for further evaluation. This process of model ranking revealed the influence of each variable in the models and showed that variables have varying significance between seasons. Leave-one-out cross validation (LOOCV) was performed as an additional method to test model skill. Results showed that all models within delta AIC of less than or equal to 2 have similar predictive skill, and taking an additional step of screening predictors leads to increased model accuracy. While statistics imply that predictor selection significantly affect the models, future projections are needed as a next step to quantify the differences in projected rainfall resulting from changes in the selection of predictors.